Indic Translate
Indic Translate is a dedicated translation model derived from google/gemma-4-E4B-it. It supports 44 English–Indic directions, Romanized text, transliteration and code-mixed translation.
Restricted access
Its prompt contract is narrower than a chat assistant
The card specifies one user message with the target language, no system message, and the supplied chat template. It recommends greedy decoding. Some long Romanized documents can end incomplete. The model is not trained to answer questions or summarize the supplied document.Indic Translate official model card
Keep translation instructions separate from general assistance
Practical guidance. Use this model when the requested output is a translation. A workflow that also asks questions about a document needs a separate step. Review number formats and untranslated code comments in technical documents.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Restricted access |
| Access | Downloadable, API |
| Model identifiers | indic-translateindic-transliterate |
| Context window | 32,768 tokens Validated document-translation length, not the larger architectural maximum. Indic Translate official model card |
| Parameters | 7.94 billion.Indic Translate official model card |
| Training | About 28.7 billion tokens, with roughly 90% core translation data and 10% Romanized translation or transliteration data.Indic Translate official model card |
| Context scope | 131,072 tokens in the architecture; document translation validated to 32,768 tokens.Indic Translate official model card |
| Access and terms | Gated weights under Indic Open Model License v1.0; upstream terms also apply.Indic Open Model License v1.0 |
| Hosted access | Bodhan documents an authenticated API operation for this model.Bodhan API reference |
Related models
Document recognition system combining trained layout and text models to produce reading-ordered Markdown and block-level JSON.
Launch date not confirmedRestricted accessDownloadable · API